ProtoShotXAI: Using Prototypical Few-Shot Architecture for Explainable AI
Authors: Samuel Hess, Gregory Ditzler
JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This section empirically demonstrates the proposed Proto Shot XAI approach on three benchmark datasets, namely: MNIST, Omniglot, and Image Net. Proto Shot XAI is compared against several state-of-the-art XAI approaches, including Ex Matchina, Ex Matchina*, Grad CAM, Grad-SHAP, RISE, and LIME. We describe multiple experiments where Proto Shot XAI can be used to explore the original classification network qualitatively and quantitatively, and demonstrate how our approach achieves comparable and often superior results to more complex algorithms for XAI. |
| Researcher Affiliation | Collaboration | Samuel Hess EMAIL Gregory Ditzler EMAIL Epi Sys Science Poway, CA 92064, USA. S. Hess was affiliated with the University of Arizona and G. Ditzler was affiliated with the University of Arizona and Rowan University when this work was performed. |
| Pseudocode | Yes | Algorithm 1 Implementation of Proto Shot XAI to Create Feature Attribution Maps |
| Open Source Code | Yes | Publicly available code & data at https://github.com/samuelhess/Proto Shot XAI/ which includes cross-comparisons between common XAI approaches with consistent representations as well as four unique experiments for XAI: adversarial MNIST, revolving six, Omniglot XAI, and a multi-class Image Net example. |
| Open Datasets | Yes | In this work, we use three commonly used datasets to evaluate the interpretability of our approach: MNIST, Omniglot, and Image Net. The MNIST dataset is a wellestablished benchmark for image classification. The Omniglot dataset is a character database that is similar to MNIST but is used to benchmark few-shot algorithms. ... (Lake et al., 2011). Image Net is a low-resolution image database that is also known as the ILSVRC-12 database (Russakovsky et al., 2015b). |
| Dataset Splits | Yes | In this dataset [Omniglot], 1,200 characters are used for training/validation and the remaining set of 423 characters are used for a disjoint testing set. Within their respective disjoint sets, images are rotated in multiples of 90 degrees to produce a set of 4,800 classes (96,000 samples) for training/validation and 1,692 classes (423 with 90 rotations) for testing (33,840 samples). |
| Hardware Specification | Yes | The algorithms were run on a Lambda Labs node with 56 CPUs, 400 GB RAM, and four Nvidia A6000 GPUs with 48 GB of RAM each. |
| Software Dependencies | No | The paper mentions using the ADAM optimizer and Keras applications library, but it does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | To train the MNIST and Omniglot networks, we used the ADAM optimizer (Kingma and Ba, 2015). The learning rate for the MNIST and Omniglot network was initially set to 6×10−5 and 1×10−3, respectively. A scheduler was used for the Omniglot experiments to decrease the learning rate by half every 100 epochs for a total of 1,000 epochs. ... For the ResNet50 Stylized-Image Net training, the SGD optimizer was used with momentum. The initial learning rate was set to 0.1 and decreases by 90% every 20 epochs for a total of 40 epochs. |